E: one word→program-name normalizer, pwsh positional + -Command, oracle matrix + PowerShell lanes, check 11 (F-460, closes #2)

[BradleyDB]

Base branch must be dev, not main. (It is.)

Regenerated output committed? The plugin bundle was rebuilt (build/build-plugin-gs-superadmin.mjs) and produced no drift; no generator or embedded doc changed.

What

Session E of the September round plan (GD-1, GD-2, GD-3 — issue #2 items 1–3 and 8, and check 11's items 4–7). Bus F-460 logged and FIXED on this branch (Class: second-scanner). Plugin 0.41.2.

• GD-1 · one word→program-name normalizer (hooks/gs-admin-guard.mjs). isGsAdminWord and normalizeProg were two normalizers; the assignment strip lived only in the first, so every interpreter behind a PowerShell assignment — $x=iex '…', $x=bash -c '…', $x=powershell "…", $x=cmd /c '…', and the ${x}= / $env:X= spellings — ran the mutation with no ask and no journal row (PowerShell assigns the interpreter's output; measured on 5.1 through the oracle). normalizeProg is now the one normalizer (assignment strip → last path segment → case-fold + launcher suffix) and isGsAdminWord derives from it; the pipe-safety lint's consumer lookup and the nested-interpreter table go through the same function.
  • Sibling found by the sweep and fixed in the same change: the assignment grammar's variable arm spelled the name's alphabet by hand ([A-Za-z_][\w:]*), while the computed-name branch carried a second, broader assignment regex. PowerShell runs $1=gs-admin jo p save, ${1}=…, $x.y=… and $a[0]=… (even when the property assignment then errors — measured), and all four were silent on dev. One grammar now ($-led word up to its first =), a superset of the old at the same position, so it can only add asks; the computed-name branch derives from it.
  • Item 1, with a deviation from the issue's recommended fix: the pwsh positional branch ended the payload search at the first non-option word, so powershell foo.ps1 -Command "…" never reached its flag. The issue recommended "flag path first, positional only as a fallback". Measured on 5.1 before choosing: the FIRST positional is the command text — powershell "gs-admin jo p save" -Command "echo hi" runs the mutation with -Command echo hi as arguments — so that order would have dropped a real ask (tenet 3). Both the positional and the -Command payload are re-scanned now. foo.ps1 -Command "…" runs nothing on 5.1 (and -File s.ps1 -Command "…" hands -Command to the script), so those asks are over-asks, the safe direction; the oracle's item-1 rows record it.
• GD-2 · the combination matrix lives in the oracle's generator (test/guard-oracle.mjs; ruling: Bradley, 2026-09-16, after a decision-tree walk-through). Assignment prefix ($x=, ${x}=, $env:X=, $1=, $x.y=, $a[0]=, the spaced control) × interpreter (bash/sh -c, option-after-flag, iex, iex -Command, Invoke-Expression, {ps} -Command, option-before-flag, positional, option-then-positional, positional-then--Command, cmd /c, eval) plus the item-1 rows, on the PowerShell lane; the $x= × interpreter crosses and the name-in-a-variable × payload crosses on the bash lane. Count: 51 PowerShell matrix rows per lane + 8 bash matrix rows (the PowerShell lane goes 47 → 98 rows, bash 603 → 611) (the full product is pruned where a prefix exercises only the prefix arm — the review's efficiency rule).
  • Every PowerShell on the host is a lane (ruling A): the oracle detects powershell.exe (5.1) and pwsh (7), runs the PowerShell rows under each, spells nested-PowerShell rows with the lane's own executable ({ps}), and prints per-lane version + counts. Effect: the ubuntu leg a dev PR runs now judges the PowerShell rows under pwsh 7 (every GitHub-hosted image ships it) — before, they were skipped on every leg but Windows, so no PowerShell row was ever judged on a dev PR. Windows judges both. A row one version does not execute (7 reads a bare positional as a file path) is unjudged there and counted, not assumed. cmdExe rows claim mustExecute on win32 only (the bash4 pattern).
• GD-3 · check 11 (build/check-doc-drift.mjs), each item reproduced by a fixture in build/test-check-doc-drift.mjs against the pre-change checker before the change (all five went red; log quoted in the review notes):
  • item 4 (reproduced, fixed): the pin claim's recognised spellings are a PIN_CLAIMS table (universal / split); a block that talks about pinning in an unknown wording is reported as such, pointing at the table, never as "no claim at all".
  • item 5 (reproduced, fixed — with a measurement that sharpened it): the anchor is located ONCE, from the id match, and its grammar accepts an id list wrapped across > quote lines (the old [^>] class refused a wrapped anchor outright as "no anchor"). The literal item-5 shape — a continuation line without > — did pass the first pattern and fail the second (block from index −1, every marker "missing"); that line also ends the callout in CommonMark, so it is red truthfully now, naming the block's line range.
  • item 6 (reproduced the mechanism; not a defect, closed with the measurement): a blank line inside the callout ends the block — CommonMark ends the blockquote there too, so the pin claim past it really has left the callout a user reads. The messages now name the block's line range so the truncation is seen rather than the check loosened; pinned as the decision.
  • item 7 (reproduced, fixed): the block is the whole callout (the walk runs backward to its first line as well as forward), so a marker whose sentence sits in the bullet above the anchor is found; a control pins that a marker outside the callout is still missing.
• Fixture pins for message text in test/guard-fixtures.mjs (20 new; 17 red against the pre-fix hook); README parsing-boundary bullet extended (no residual closed — the JSON set is unchanged); CHANGELOG 0.41.2.
• Review round (/code-review medium: 8 finders, 3 file-batched verifiers; 9 confirmed, 1 plausible, 7 cleanups applied, 8 dismissed). Two confirmed defects fixed here: (a) a live bypass on every shipped version — a value-taking PowerShell option before the payload (powershell -ExecutionPolicy Bypass "gs-admin jo p save", -InputFormat Text …, and behind $x=) had its VALUE read as the positional payload, so the real one was never re-scanned; the value-option table gains a pwsh predicate (value-taking parameters by prefix; valueless ones excluded so no payload can be skipped) shared by the flag loop and rescanFrom; (b) a regression this branch had introduced — with the break gone, powershell "gs-admin jo p save" -Command "gs-admin jo p save" journaled two rows for one execution; a positional that carried an invocation now ends the search. Plus: the oracle's on: lane gate (positive claims measured only on 5.1 are asserted on that lane; executed ⇒ guarded holds everywhere), matrix breadth as a data column, one duplicate row removed, isAssignmentWord named once, normalizeProg read once per word, stale "two scanners" comments corrected, workflow comment updated, check 11 on lib.mjs's QUOTE_RE with unterminated-anchor and anchor-outside-callout refusals. Dismissed with reasons: pruning the ${x}= full cross (the braces axis is the issue's own), the double PowerShell lane on Windows (ruling A), Linux-unrunnable rows counted as over-ask (that is what they are), the nested powershell spelling on pwsh-only hosts (a name the host lacks cannot be judged), a preamble marker satisfying the whole-callout block (item 7 by design), deriving the split claim from data (F-252's deliberate false-direction-only rule), merging green drift mutants, and the PowerShell start-up double spawn.

Why

Issue #2 (re-homed from the private bus's F-253): silent misses on the safety surface, measured by mutation, plus check 11's brittleness. The class is second-scanner (F-443): two readers re-deriving one grammar disagree exactly where it is subtle — here, which words are assignments and what a word's program name is.

Review notes

Oracle counts (the judge of record — test/guard-oracle.mjs, this Windows host: bash 5.x + Windows PowerShell 5.1; no pwsh on this machine, so the 7 lane's first measurement is this PR's ubuntu suite):

• Pre-fix hook (dev at PR D: --deep and Phase 6 boundaries (F-457), guard repeat-ask remedy (F-453), work-ordering bullet (#1) — plugin 0.41.1 #22's merge, run through GUARD_ORACLE_HOOK against a git archive of the dev plugin tree — a lone hook file fails open without its bundled catalog, which the first attempt proved): 709 generated lines run (611 bash 5.x; 98 PowerShell rows × 1 lane(s) — powershell 5.1.26100.9444: 98 rows, 44 bypass, 3 over-ask); 44 bypass, 52 over-ask, 44 over-qualified, 9 documented residuals held — 80 failures; every bypass is a matrix row (14 under $x= and 14 under ${x}= — the 12 assignment × interpreter crosses each plus the two value-option crosses — 3 under $env:X=, 3+1 under $1=, 3+1 under $x.y=, 3+1 under $a[0]=, 1 under ${1}=).
• Fixed hook: 709 generated lines run (611 bash 5.x; 98 PowerShell rows × 1 lane(s) — powershell 5.1.26100.9444: 98 rows, 0 bypass, 7 over-ask); 0 bypass, 62 over-ask, 44 over-qualified, 9 documented residuals held — all checks passed.
• The over-ask delta is exactly the non-executing shapes the generator added (bash $x= forms, $x=eval, the two item-1 scripts).

Pre-change drift-test run (build/test-check-doc-drift.mjs with build/check-doc-drift.mjs from dev): 11b red "makes no claim about pinning at all" (item 4); 11c wrapped anchor red "carries no anchor"; 11c literal shape red with every marker missing; 11d red (same as after, the message differs); 11e red "readmeMarker no longer appears … zzz-above-anchor" (item 7). After: 129 checks green (11f, the unterminated-anchor refusal, added by the review round).

Guard-fixtures against the pre-fix hook: 17 red (10 interpreter/positional pins, 4 assignment-sibling pins, 3 value-option pins; the single-row pin is green there because the old break kept one row), controls green; after: 697 green.

Where to look first: the one normalizer and ASSIGNMENT_PREFIX (hook ~lines 240–300); the pwsh positional loop (~1450); the oracle's lane detection and matrix (~110, ~545); check 11 (~1286–1400).

Battery run locally, verbatim: guard-fixtures, guard-oracle, doc-lib-fixtures, contract-conformance, test-check-doc-drift, test-check-stale-facts, test-defect-classes, check-stale-facts, check-doc-drift, check-imports, check-instance-data, npm run typecheck, plugin rebuild + git diff (no drift), claude plugin validate ×2 (strict).

Issue #2 disposition: items 1, 2, 3, 8 closed by this PR; items 4, 5, 7 fixed with fixtures; item 6 measured not a defect (pinned as the decision). All eight addressed → Closes #2 — the E-V verdict comments the oracle's counts on the issue.

• Plugin version bumped in .claude-plugin/plugin.json with a CHANGELOG entry (0.41.2 — the hook's decisions change).
• Safety boundary: this PR touches hooks/gs-admin-guard.mjs. No "ask" became a "deny" or disappeared: the normalizer change is a superset match (every word that matched before matches now with the same remainder), the assignment grammar is a superset at the same position, and the positional branch re-scans MORE payloads, never fewer. test/guard-fixtures.mjs extended (16 pins) and the oracle's generator extended (the matrix + lanes).
• Skill edits: none (no SKILL.md changed; README prose only).
• No test or dev step touched real gs-admin state.
• No tenant/instance/org-specific data in commit messages or this PR body.

Closes #2

[BradleyDB]

Dispatch evidence on the handoff tip 58b1330 (hb-20260916-02), quoted after completion: validate-plugin 35147717941 (PR) changes/validate (ubuntu)/manifests all success; docs-drift 35147717956 (PR) drift (full) success; dispatches 35147785798 (validate-plugin, all jobs success) and 35147788662 (docs-drift, full + stripped success). The pwsh 7 lane's first measurement, from the ubuntu job log: 709 generated lines run (611 bash 5.x; 98 PowerShell rows × 1 lane(s) — pwsh 7.6.5: 98 rows, 0 bypass, 20 over-ask); 0 bypass, 78 over-ask, 44 over-qualified, 9 documented residual(s) held — All guard-oracle checks passed. The two PowerShell residual rows executed under 7 too; the 13 extra over-asks are the positional shapes 7 reads as a file path.